LangMolDiode Checkpoints
Collection
SFT and reinforcement-learning checkpoints for LangMolDiode. Project code: https://github.com/feiyang-cai/LangMolDiode • 3 items • Updated
LangMolDiode Qwen3.5-4B RL LoRA
LangMolDiode is a language-conditional molecule generator that generates SMILES from detailed molecular structure descriptions.
Project code: LangMolDiode
This repository contains adapters from the LangMolDiode Qwen3.5-4B DAPO run. The root adapter is the final rollout step 1500. Additional selected adapters are available under:
| Subfolder | Selection set | Rollout step | exact@1 |
|---|---|---|---|
best_validated_step1260/ |
1,972 validated prompts | 1260 | 0.8519 |
best_test_step1315/ |
200 MolLangBench test prompts | 1315 | 0.7150 |
best_extended_step1370/ |
200 MolLangBench extended prompts | 1370 | 0.7750 |
Each validation prompt was sampled three times. exact@1 is the first sample's
canonical-SMILES exact match. Best means highest observed exact@1 among the
full saved validation runs, with the earliest step winning a tie. Selection on
the test and extended sets means their reported values should be treated as
selection results, not untouched holdout estimates.
The adapters use the merged SFT base ChemFM/LangMolDiode-Qwen3.5-4B-SFT. LoRA rank is 64 and alpha is 128.
- Downloads last month
- 15
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support
Model tree for ChemFM/LangMolDiode-Qwen3.5-4B-RL-LoRA
Base model
Qwen/Qwen3.5-4B-Base Finetuned
Qwen/Qwen3.5-4B Finetuned
ChemFM/LangMolDiode-Qwen3.5-4B-SFT